Former OpenAI Researchers Challenge Dismissals

Three former safety researchers at OpenAI have released an open letter contesting the company’s reasons for their dismissal. They expressed concern that the incident might discourage other employees from openly discussing AI risks.

Details of the Open Letter

Tomek Korbak, Jasmine Wang, and Mikita Balesni addressed their letter to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. The researchers claim communications around their firing have left former colleagues feeling “afraid to speak.” They emphasized the importance of challenging decisions and collaborating with external experts in AI safety. They noted, “AI is not a normal technology, and OpenAI is not a normal company.”

OpenAI refuted the suggestion that the dismissals were linked to safety concerns. The company cited a “thorough investigation” that determined policy violations involving sensitive information.

Dispute Over Dismissal Reasons

In their letter, titled “OpenAI cannot make AI safe on its own,” the researchers expand on the reasons for their firing. They opened with an assertion about the impact of their dismissal on other employees’ willingness to speak. They claim staff are now “unclear on where they stand.” The trio highlight the importance of a safe working environment for addressing AI safety concerns without fear.

Personal Accounts from the Researchers

Korbak, Wang, and Balesni dispute claims of leaking AI architecture information. They deny improper communication with external organizations, asserting their interactions were within job requirements.

Korbak mentions he was verbally informed that his dismissal involved communication with METR, a group evaluating AI systems. He shared that no written details were provided, stating, “Talking to METR was my job.” He also raised concerns about a decreasing ability to monitor AI agents’ reasoning, believing this led to his dismissal.

Balesni challenges OpenAI’s reasoning further, claiming prioritizing safety over corporate interests led to their firing. He asserts that interactions with third-party safety organizations were coordinated with his direct superiors, removing sensitive details as required.

Wang disputes claims that accessing an executive’s email led to her dismissal. She maintains the access was delegated for recruiting purposes and disclosure of any accidental oversight was immediate.

Researchers’ Recommendations

The dismissed researchers propose several actions for OpenAI and other AI companies. They recommend continued independent safety evaluations, protection for monitoring systems, and maintaining open communication with safety experts. Balesni expressed concern about the firings impacting relations with METR.

OpenAI’s Response

On October 9, OpenAI released a statement defending the dismissals, citing breaches of trust and policy violations. It denied that speaking out on safety led to the firings, affirming continued encouragement of safety debates within the company.

The company emphasized its commitment to independent safety assessments, noting it is finalizing contracts with external evaluators. OpenAI affirmed agreement with the researchers on the importance of monitoring advanced models, pointing to ongoing research efforts in this area.

OpenAI acknowledged the researchers’ contributions to AI safety, yet stood firm on the dismissal decision due to alleged breaches.

Conclusion

The disagreement underscores the tension in managing sensitive information while engaging with external safety evaluators. Both parties agree on principles such as independent evaluation but disagree on the events and policy clarity leading to the dismissals.